Supermix v103 (v103_scf_synthesis.pt β€” Engram + Gated DeltaNet-2 + Frontier Synthesis)

Standalone flagship experimental release for v103_scf_synthesis.pt and the complete v103 Frontier Synthesis stack from kai9987kai/Supermix (main commit db535a42).

What v103 Adds

  1. Full-Stack Frontier Neural Checkpoint (weights/v103_scf_synthesis.pt):
    • Warm-started from Kai9987kai/supermix-v102 (v102_scf_frontier.pt, 9,460-token vocab; 37,561,559 total / 10,094,663 active params) using MuonHybridOptimizer.
    • Adds Engram Conditional Memory (use_engram_memory=True, engram_table_size=2048, engram_n_heads=2, engram_after_layer=0): DeepSeek-2026-style O(1) multi-head hashed causal 2-gram/3-gram table lookup with context-aware gating and rank-3 tail-cache rollback parity.
    • Adds Gated DeltaNet-2 Fast-Weight Memory (use_deltanet_memory=True, deltanet_dim=32, deltanet_after_layer=1): decoupled channel-wise decay, erase, and write linear-time matrix memory recurrence with exact KV-cache and MTP speculative rollback parity.
    • Composes seamlessly with ManifoldHyperConnection (mHC), ConnectomeCore (dual_soma + commissural_gate + glial_homeostasis + hebbian_trace + predictive_error + divisive_norm + dendritic_apical + synaptic_homeostasis_sleep_cycle), and RecursiveThinkingCore (entropy_halt + ctm_synchrony).
    • Achieves lowest seen-response loss in lineage (tier1_seen_response = 0.0047, ppl 1.0047), dev_loss = 0.0180 (ppl 1.02), tier3 = 0.0515 (tier3/tier1 = 1.048x), connectome ablation +0.00007 nats, and all 5/5 v58 generalisation checks [PASS] including exact speculative_matches_greedy.
  2. Five Frontier Synthesis Diagnostic Engines (source/nexus_v103_synthesis.py, all answer_authority=False with SHA-256 receipts):
    • ContinuumBTSPHopfieldMemory (+ Poseidon MNEMOSYNE 5-mode lesion assay), TopKSparseAutoencoderProbe (overcomplete Top-K SAE + causal feature ablation), FractalConnectomeAnalyzer (Sierpinski / Hilbert / Small-World / Random Control), CausewaySensitivityWorkbench, and NeuroHiveResourcePlanner.
  3. Exact Linear-Algebra & Series Solvers + Fail-Closed Studio Frontier Lab:
    • Exact Fraction 2x2 matrix inverse, 2x2 real eigenvalues, dot/cross products, finite/infinite geometric series, and polynomial evaluation; /v1/v103/* FastAPI endpoints; and web_static/nexus_studio.html (#panel-v103-frontier).

Honest Measured Benchmarks (Windows CPU, max_new_tokens=96)

Evaluation Suite Result Receipt
Generalisation Ladder (v58) dev_loss = 0.0180 (tier1 = 0.0047, tier3 = 0.0515, 5/5 PASS) results/generalisation_results.json
Standalone Neural seed=82 (41 tasks) 34 / 41 (0.8293) results/eval_v103_seed82.json
Standalone Neural seed=83 (82 tasks) 71 / 82 (0.8659) results/eval_v103_n82_seed83.json
Standalone Neural seed=84 (82 tasks) 72 / 82 (0.8780) results/eval_v103_n82_seed84.json
3-Seed Combined Standalone (205 tasks) 177 / 205 (86.3%) Paired against v96 (180/205), v100 (179/205), v101 (178/205)
SCF Hybrid (python -m scf bench --seed 82 --neural) 41 / 41 (1.0000) results/scf_bench_v103_seed82.json

Per Rule 9 promotion gating (automatic_promotion_enabled=False), v96_physio_neurogenesis.selected.pt (Kai9987kai/supermix-v96) remains #1 Promoted Default; v103_scf_synthesis.pt is released as Experimental.

Quick Start

pip install torch numpy flask fastapi uvicorn huggingface_hub
huggingface-cli download Kai9987kai/supermix-v103 --local-dir .
python example_usage.py
PYTHONPATH=source python -m scf frontier --json
python source/nexus_api.py --host 127.0.0.1 --port 8000        # Studio: http://127.0.0.1:8000/studio
python source/supermix_chat_server.py --model v103_scf_synthesis=weights/v103_scf_synthesis.pt --port 8780
Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for Kai9987kai/supermix-v103